Embed Size (px)
Citation preview
IMPROVEMENT OF AUDIO FEATURE EXTRACTION TECHNIQUES IN
TRADITIONAL INDIAN STRING MUSICAL INSTRUMENT
KOHSHELAN A/L SUNDARARAJOO
A thesis submitted in
fulfillment of the requirement for the award of the
Degree of Master of Information Technology
Faculty of Computer Science and Information Technology
Universiti Tun Hussein Onn Malaysia
AUGUST 2015
vi
ABSTRACT
Audio feature extraction is an essential and significant process where audio features are
extracted from the audio files whereby the extracted audio features contains
relevant audio information. One of the important roles played by the audio features is to
improve the classification accuracy. However, the presence of noise in the audio signals
which degrades the quality of the extracted features may result in low classification
accuracy. Some of the existing audio feature extraction techniques are Mel-Frequency
Cepstral Coefficient (MFCC), Linear Predictive Coding (LPC), Local Discriminant
Bases (LDB), Zero-Crossing Rate (ZCR) and Perceptual Linear Prediction (PLP).
Furthermore, the three frequently used techniques in audio feature extraction are MFCC,
LPC and ZCR. Previous research had mentioned the shortcomings of the three
techniques on extracting noisy signal. This has been identified in the case of traditional
Indian musical instrument where the vibration of string instrument had produced noise
in the highest amplitude. Therefore, Zero Forcing Equalizer (ZFE) was proposed to
equalize the noise in the highest amplitude. ZFE was integrated with three audio feature
extraction techniques, namely MFCC-ZFE, LPC-ZFE and ZCR-ZFE in order to improve
the performance of the existing techniques. The results show the best improvement of
classification accuracies obtained for the proposed techniques of MFCC-ZFE were
98.2% of classification accuracies with 4.0% of improvement by using kNN. Meanwhile,
the combined features of the MFCC-ZFE + LPC-ZFE + ZCR-ZFE have obtained 98.3%
of classification accuracies with 9.1% of improvement by using kNN.
vii
ABSTRAK
Pengekstrakan ciri audio adalah satu proses yang penting di mana ciri-ciri audio yang
diekstrak dari fail audio mengandungi maklumat yang relevan. Salah satu peranan
penting yang dimainkan oleh ciri-ciri audio adalah untuk meningkatkan prestasi
pengkelasan audio. Walau bagaimanapun, disebabkan oleh kehadiran bunyi hingar di
dalam isyarat audio mengurangkan kualiti ciri-ciri yang diekstrak serta merendahkan
prestasi pengkelasan audio. Beberapa teknik pengekstrakan ciri audio yang sedia ada
ialah Mel-Frequency Cepstral Coefficient (MFCC), Linear Predictive Coding (LPC),
Local Discriminant Bases (LDB), Zero-Crossing Rate (ZCR) dan Perceptual Linear
Prediction (PLP), tetapi tiga teknik yang sering digunakan ialah MFCC, LPC dan ZCR.
Kajian sebelum ini telah mengenal pasti kelemahan dalam tiga teknik tersebut yang
disebabkan oleh getaran daripada alat muzik tradisional India khususnya alat muzik
bertali yang menghasilkan bunyi hingar dalam amplitud tertinggi. Oleh itu, Zero Forcing
Equalizer (ZFE) dicadangkan untuk menyelaraskan bunyi hingar dalam amplitud
tertinggi. Dalam kajian ini, ZFE diintegrasikan dengan tiga teknik pengekstrakan ciri
audio, iaitu MFCC-ZFE, LPC-ZFE dan ZCR-ZFE untuk meningkatkan prestasi teknik
yang sedia ada. Hasil penelitian menunjukkan peningkatan ketepatan pengkelasan audio
yang terbaik diperolehi oleh MFCC-ZFE ialah 98.2% pengkelasan audio dengan 4.0%
peningkatan ketepatan pengkelasan audio menggunakan kNN. Selain itu, penggabungan
ciri-ciri audio daripada MFCC-ZFE + LPC-ZFE + ZCR-ZFE memperoleh 98.3%
pengkelasan audio dengan 9.1% peningkatan ketepatan pengkelasan audio menggunakan
kNN.
viii
CONTENTS
TITLE i
DECLARATION ii
DEDICATION iii
ACKNOWLEDGEMENT iv
PUBLICATION v
ABSTRACT vi
ABSTRAK vii
CONTENTS viii
LIST OF TABLES xii
LIST OF FIGURES xiv
LIST OF SYMBOLS AND ABBREVIATIONS xv
LIST OF APPENDICES xvi
CHAPTER 1 INTRODUCTION
1.1 Background of Study 1
1.2 Research Motivation 3
1.3 Objectives 5
1.4 Scope of Research 5
1.5 Importance of Research 6
1.6 Thesis Outlines 6
ix
CHAPTER 2 LITERATURE REVIEW
2.1 Introduction 8
2.2 Audio Feature Extraction Concept 8
2.3 An Overview of Audio
Feature Extraction Techniques 9
2.3.1 Mel Frequency Cepstral Coefficients
(MFCCs) 13
2.3.2 Linear Predictive Coding (LPC) 16
2.3.3 Zero-Crossing Rate (ZCR) 18
2.4 Noise in Audio Signal 19
2.5 Noise Filtering 20
2.6 Sound Equalizer 20
2.6.1 Zero Forcing Equalizer (ZFE) 22
2.6.2 Minimum Mean Square Error
(MMSE) Equalizer 24
2.7 Audio Classification Techniques Overview 25
2.7.1 K-Nearest Neighbor (kNN)
Classifier 25
2.7.2 Bayesian Network (BNs)
Classifier 26
2.7.3 Support Vector Machine (SVM)
Classifier 27
2.7.4 C4.5 Decision Tree Classifier 28
2.7.5 Naïve Bayes (NB) Classifier 29
2.8 An Overview of Instrument Sounds 30
2.9 Traditional Indian Musical Instrument 30
2.10 String Instrument 31
2.11 Carnatic music 32
2.12 Veena 33
2.13 Chapter Summary 35
x
CHAPTER 3 METHODOLOGY
3.1 Introduction 36
3.2 Research Methodology 36
3.2.1 Audio Acquisition 38
3.2.2 Audio Pre-processing 39
3.2.3 Audio Feature Extraction 40
3.2.3.1 The Proposed Technique of
MFCC-ZFE 42
3.2.3.2 The Proposed Technique of
LPC-ZFE 45
3.2.3.3 The Proposed Technique of
ZCR-ZFE 47
3.2.4 Audio Classification 50
3.2.5 Performance Measurement 50
3.3 Chapter Summary 52
CHAPTER 4 ANALYSIS AND RESULT
4.1 Introduction 53
4.2 Performance of Audio Feature Extraction
Techniques 53
4.3 ZFE and MMSE Equalizer Performance
Comparison 59
4.4 Chapter Summary 66
CHAPTER 5 CONCLUSION
5.1 Research Summary 67
5.2 Research Contributions 68
5.3 Research Advantages 69
5.4 Future Works 69
5.5 Conclusion 70
xi
REFERENCES 71
APPENDIX 82
xii
LIST OF TABLES
2.1 Comparison of audio feature extraction techniques
used by authors 10
2.2 Descriptions of five instruments from the class
of Veena 34
3.1 Audio Files Properties 39
3.2 Experimental Parameter for MFCC, LPC and
ZCR 41
3.3 The original Pre-emphasis algorithm 44
3.4 The modified Pre-emphasis algorithm 44
3.5 The original LPC Synthesizer algorithm 46
3.6 The modified LPC Synthesizer algorithm 46
3.7 The original ZCR energy calculation algorithm 48
3.8 The modified ZCR energy calculation algorithm 49
4.1 A comparison of the audio classification accuracy
between the original and the proposed techniques
in five different classifiers 54
4.2 A comparison of the audio classification accuracy
by combining the audio features in five
different classifiers 56
4.3 A comparison between classification accuracy and
RMSE for the proposed techniques 57
xiii
4.4 A comparison between classification accuracy,
F-Measure and MAE for the proposed techniques 58
4.5 A comparison between the ZFE and MMSE
equalizer 59
4.6 Confusion matrices for MFCC in five different
classifiers 60
4.7 Confusion matrices for MFCC-ZFE in five
different classifiers 61
4.8 Confusion matrices for LPC in five different
classifiers 62
4.9 Confusion matrices for LPC-ZFE in five
different classifiers 63
4.10 Confusion matrices for ZCR in five different
classifiers 64
4.11 Confusion matrices for ZCR-ZFE in five
different classifiers 65
xiv
LIST OF FIGURES
2.1 MFCC Block diagram 13
2.2 Spectrogram diagram 14
2.3 LPC Block diagram 17
2.4 ZCR Block diagram 18
2.5 Audio equalizer 21
2.6 Graphical model of BNs 27
2.7 Hierarchical of Traditional Indian Musical
Instrument 31
3.1 Research Methodology 37
3.2 Research Framework 37
3.3 Audio signal is cropped in 5 seconds 40
3.4 Spectrogram 42
3.5 Modified Pre-emphasis block diagram 43
3.6 Spectrogram for original MFCC 45
3.7 Spectrogram for MFCC-ZFE 45
3.8 Modified LPC Synthesizer block diagram 45
3.9 Spectrogram for original LPC 47
3.10 Spectrogram for original LPC-ZFE 47
3.11 Modified ZCR energy calculation block diagram 48
3.12 Spectrogram for original ZCR 49
3.13 Spectrogram for original ZCR-ZFE 49
xv
LIST OF SYMBOLS AND ABBREVIATIONS
f - Frequency
dB - Decibel
KHz - Kilohertz
MFCC - Mel Frequency Cepstral Coefficient
LPC - Linear Predictive Coding
ZCR - Zero Crossing Rate
kNN - k-Nearest Neighbor
BNs - Bayesian Network
SVM - Support Vector Machine
SMO - Sequential Minimal Optimization
NB - Naïve Bayes
ISI - Inter-Symbol Interference
ZFE - Zero Forcing Equalizer
MIMO - Multiple Input Multiple Output
MLSE - Maximum Likelihood Sequence Estimator
CBMIR - Content-based Music Information Retrieval
FFT - Fast Fourier Transform
RMSE - Root Mean Square Error
MAE - Mean Absolute Error
xvi
LIST OF APPENDICES
APPENDIX TITLE PAGE
A Classification averages for original and
proposed techniques 82
B Confusion matrices for combined features
in five different classifiers 84
CHAPTER 1
INTRODUCTION
1.1 Background of study
Audio feature extraction is a process to obtain a set of audio features from an audio signal
whereby the rich set of audio features can be utilized in the computation aspect such as in
determining the average, maximum or frequency values that can be plotted to a spectrogram
(Dave, 2013; Bullock, 2008). There are two types of extracted features that are low-level features
and high-level features (Lerch, 2012; Furht, 2009). Low-level features represent terms in which
humans refer to music such as pitch, tempo, amplitude and others. High-level features are
considered to have direct (humanly interpretable) and derived from the low-level features such as
genre and style. This research will be focusing on the low-level features since the research will
extract the audio features from the high amplitude of the signal. Audio feature extraction can
contribute for better audio classification accuracy result depending on the quality of the extracted
features (Umapathy, Ghoraani & Krishnan, 2010). However, one of the essential drawbacks in
the audio feature extraction process is the presence of disturbance such as noise from the high
amplitude signal which is produced from certain instrument that may degrade the quality of the
extracted features and lead to low classification accuracy (Stulov & Kartofelev, 2014; Dave,
2013; Wolf & Nadeu, 2008; Subramanian, 2006). There are many techniques for audio feature
2
extraction, such as Mel-Frequency Cepstral Coefficients (MFCC), Local Discriminant Bases
(LDB), Linear Predictive Coding (LPC), Perceptual Linear Prediction (PLP) and Zero-Crossing
Rate (ZCR) (Dave, 2013; Anusuya & Katti, 2011; Umapathy, Krishnan & Rao, 2007). However,
three audio feature extraction techniques are selected which are Mel-Frequency Cepstral
Coefficients (MFCC), Linear Predictive Coding (LPC) and Zero-Crossing Rate (ZCR) based on
their good performance and frequently used by previous researchers (Chougule & Chavan, 2014;
Bormane & Dusane, 2013; Anusuya & Katti, 2011; Kumari, Kumar & Solanki, 2010;
Gunasekaran & Revathy, 2008).
MFCC is one of the techniques commonly used in digital signal processing (Weeks,
2010). MFCC has been proven as one of the effective techniques in audio feature extraction
(Keronen et al., 2011; Furht, 2009). Similar to MFCC, LPC is another technique which offers a
powerful, yet simple method to extract the audio information (Elminir, ElSoud & El-Maged,
2012; Sheetal & Raut, 2012). On the other hand, ZCR is useful for musical instruments
measurement and endpoint detection (detection of the start and end of unvoiced sounds) (Khan,
Bhaiya & Banchhor, 2012). Even though these three techniques have performed well in
extracting audio features, they have shown some shortcomings in extracting noisy signal
especially from string instrument in the domain of traditional Indian musical instrument
(Anusuya & Katti, 2011; Chougule & Chavan, 2014; Bormane & Dusane, 2013).
Traditional Indian musical instrument is one of the oldest musical instruments in the
world (The Incredible India Travel, 2011). It has contributed immensely in making Indian music
famous around the world. Traditional Indian musical instrument can be categorized into four
types such as string instruments, percussion instruments, wind-blown instruments and solid
instruments (Gunasekaran & Revathy, 2008). Specifically, in the context of content-based audio
extraction and audio classification, these instruments have been used many times to identify the
characteristics of a single instrument such as pitch or amplitude among multi-instruments
(Gunasekaran & Revathy, 2008). However, it is difficult to identify the characteristics of a
particular instrument such as string instrument since each instrument creates a sound which
produces music in its own ways. Furthermore, these instruments may contribute to noisy signal
due to some disturbance from the indoor, outdoor or instrument itself. Among them, string
instruments such as Veena contributes more to noise due to sound produced from the vibration of
strings (Dave, 2013; Stulov & Kartofelev, 2014; Subramanian, 2006). In this research, string
3
instrument of traditional Indian musical instrument were used since they contribute to more noise
that come from the highest amplitude of the audio signal and lead to low quality of extracted
audio features.
Therefore, Zero Forcing Equalizer (ZFE) was proposed in this research to be integrated
with three audio feature extraction techniques, namely MFCC-ZFE, LPC-ZFE and ZCR-ZFE to
equalize the noise which comes from the highest amplitude. ZFE is a linear receiver (also known
as an equalizer) used in communication systems (Kaur & Kansal, 2013). The function of ZFE is
to invert the frequency response of the channel. ZFE has an ability to equalize the noise which
comes from the highest amplitude of the audio signals (Kaur & Kansal, 2013). ZFE is integrated
with audio feature extraction techniques to overcome the shortcoming occur in the existing
techniques as well as to improve the quality of extracted features. Furthermore, another linear
equalizer known as Minimum Mean Square Error (MMSE) equalizer is also integrated with the
three audio feature extraction techniques for performance comparison.
A good audio extraction technique will lead to better accuracy of audio classification. In
order to evaluate the performance of the proposed audio feature extraction techniques, five
benchmark classifiers were selected. These classifiers were K-Nearest Neighbor (kNN),
Bayesian Network (BNs), Support Vector Machine (SVM), C4.5 decision tree and Naïve Bayes
(NB) classifiers. The classifiers were selected based on their good performance in audio
classification (Kumar, Pandya & Jawahar, 2014; Bello, 2013; Rocha, Panda & Paiva, 2013;
Witten, Frank & Hall, 2011; Nettleton, Orriols-Puig & Fornells, 2010; Li, Ogihara & Li, 2003).
The classification results are compared to show the performance of the extracted audio features.
1.2 Research Motivation
As the need of reliable information from the audio and music grows, the importance of research
on audio feature extraction increases. Audio feature extraction has contributed immensely in
various fields such as in data mining involving Content-Based Music Information Retrieval
(CBMIR). CBMIR has become a critical research topic and has been given increasing attention
in recent years due to the extensive growth in audio and music (Yu et al., 2013). CBMIR
4
generally involves analyzing, searching and retrieving music based on audio features of an audio
which is normally used to represent songs or music genre. Identifying them would normally
involve feature extraction and classification tasks. Theoretically, the greater the features
analyzed, the better the classification accuracy can be achieved.
The impact of audio feature extraction in audio classification is huge since the
performance of audio classification accuracy can be defined based on the quality of extracted
audio features. Furthermore, good quality of extracted audio features may contribute to a better
accuracy of audio classification (Umapathy, Ghoraani & Krishnan, 2010). However, the quality
of audio features depends upon the behavior of the audio domain. Audio domain, such as
traditional India musical instrument is one of the oldest musical instruments in the world,
however, there is not much work done in the area of feature extraction compared to Western
music (Agarwal, Karnick & Raj, 2013). In the previous research, traditional Indian musical
instrument involving string instruments had shown fluctuating behavior in its audio signal during
different experimental setups that were identified by using MFCC, LPC and ZCR (Gunasekaran
& Revathy, 2008; Chougule & Chavan, 2014). This fluctuating behavior was identified due to
the vibration of string instruments which produced unwanted sound (noise) in highest amplitude
(Dave, 2013; Stulov & Kartofelev, 2014; Subramanian, 2006). Wolf and Nadeu (2008) also said
that MFCC performance degrades severely when the extracted features contain noise. Moreover,
Anusuya and Katti (2011) mentioned that if the audio signal used is noisy, the extracted features
from MFCC, LPC and ZCR lead to lower classification accuracy.
Based on the previous researches, their findings showed that the existing audio feature
extraction techniques are unable to produce high quality audio features due to the presence of
noise in the highest amplitude of the audio signal (Stulov & Kartofelev, 2014; Dave, 2013; Wolf
& Nadeu, 2008; Subramanian, 2006). This shortcoming have inspired the use of ZFE to equalize
the noise in the highest amplitude (Kaur & Kansal, 2013). Therefore, in this research, ZFE was
integrated with MFCC, LPC and ZCR to overcome the drawback of audio feature extraction
techniques in extracting noisy audio signal in highest amplitude. This research perceived that
improvement can be done to the audio feature extraction techniques to overcome the drawback
and lead to a better quality of extracted audio features.
5
1.3 Objectives
This study embarks on the following objectives:
i. To propose three (3) improved techniques of Mel-Frequency Cepstral Coefficient
(MFCC), Linear Predictive Coding (LPC) and Zero-Crossing Rate (ZCR) by integrating
them with Zero Forcing Equalizer (ZFE);
ii. To implement the proposed techniques in (i) in traditional Indian string musical
instrument; and
iii. To evaluate the performance of the proposed audio feature extraction techniques based on
audio classification accuracy by using five classifiers which are k-Nearest Neighbor
(kNN), Bayesian Network (BNs), Support Vector Machine (SVM), C4.5 decision tree
and Naïve Bayes (NB).
1.4 Scope of Research
The scope of research is divided into four parts which are audio acquisition, audio pre-
processing, audio extraction and audio classification. Audio acquisition, is the process of
collecting information on audio files such as sampling rate and bit depth. The dataset of audio
files from traditional Indian string musical instrument were collected from SoundCloud, the
online music sharing platform (SoundCloud, 2007). There were a total of 500 audio files since
each instrument from the class of Veena contribute to 100 audio files. Audio pre-processing is
the process to modify the audio signal based on user preferences. The audio files were cropped
when the amplitude value is bigger than or closer to the maximum value which is 0dB (Zytrax,
2014). Meanwhile, audio feature extraction involves the process of extracting low-level features
of audio files (Lerch, 2012; Furht, 2009). Three audio feature extraction techniques were used
which are Mel-Frequency Cepstral Coefficient (MFCC), Linear Predictive Coding (LPC) and
Zero-Crossing Rate (ZCR). Audio classification, on the other hand, is the process to categorize
the audio features into a sample of classes in order to obtain their classification accuracy by
6
using different classifiers. Five audio classifiers were used which are k-Nearest Neighbor (kNN),
Bayesian Network (BNs), Support Vector Machine (SVM), C4.5 decision tree and Naïve Bayes
(NB). In addition, in order to validate the performance of ZFE equalizer, another similar
equalizer known as Minimum Mean Square Error (MMSE) equalizer was used in this research
(Kumar and Kaur, 2012). MMSE equalizer was integrated with the audio feature extraction
techniques in the same way ZFE was integrated with the techniques. The performance of ZFE
and MMSE will be compared.
1.5 Importance of Research
The proposed techniques are important because it can be utilized in many areas or field involving
soft computing. By studying the techniques used in the audio feature extraction process, the
proposed techniques are hoped to provide a better performance of audio classification accuracy
when extracting audio signal that contains noise in the highest amplitude. Therefore, this
research hopes to solve the problem of audio feature extraction techniques in the domain of
Indian traditional instruments.
1.6 Thesis Outlines
For the remaining chapters of the thesis is structured as follows:
Chapter 2 provides a detail explanation of the audio feature extraction techniques used in the
research which is Mel-Frequency Cepstral Coefficients (MFCC), Linear Predictive Coding
(LPC) and Zero-Crossing Rate (ZCR). Moreover, this chapter also discusses on the sound
equalizer which is Zero Forcing Equalizer (ZFE) and Minimum Mean Square Error (MMSE).
The benchmark classifiers which are k-Nearest Neighbor (kNN), Bayesian Network (BNs),
Support Vector Machine (SVM), C4.5 decision tree and Naïve Bayes (NB) are used to evaluate
7
the performance of audio feature extraction techniques. In addition, the background of traditional
Indian musical instrument is mention in this chapter.
Chapter 3 outlines the phases involved in the methodology. The methodology consisted of four
phases which is audio acquisition, audio pre-processing, proposed audio feature extraction
techniques and audio classification. Besides, this chapter also provides an explanation on the
performance measurement in term of classification accuracy, Root Mean Square Error (RMSE)
F-Measure and Mean Absolute Error (MAE).
Chapter 4 presents the performance evaluation results based on the classification accuracy,
RMSE, F-Measure and MAE of the extracted audio features of audio feature extraction
techniques. In this chapter, the comparisons between the original and proposed techniques were
mentioned. In addition, a comparison between ZFE and MMSE equalizers is presented. A detail
explanation on the confusion matrices is provided.
Chapter 5 concludes and highlights the research finding. This chapter summarizes the research
outcome and discusses the advantages of the research. Besides, this chapter provides the future
work that can be done to further enhance the research.
CHAPTER 2
LITERATURE REVIEW
2.1 Introduction
This chapter outlines the techniques used in audio feature extraction and audio classification.
In addition, a detail description of the techniques were provided. Furthermore, this chapter
also provides an explanation on the sound equalizer used in the research. Besides, the
overview on the domain of tradition Indian string musical instrument was introduced in this
chapter.
2.2 Audio Feature Extraction Concept
Audio feature extraction is a process that involves transforming audio data to a set of features
such as pitch, timbre, and others (Bormane & Dusane, 2013; Bullock, 2008). Specifically,
audio feature extraction process addresses the analysis and extraction of meaningful
information from audio signals. The objective of the audio feature extraction process is to
capture the relevant information on audio signal to get a higher-level understanding of the
audio signal (Dave, 2013; Bullock, 2008). Furthermore, the extracted features of the audio
signal may provide a higher-level understanding of the amplitude or frequency components
9
of the audio signal by plotting to the spectrogram (Dave, 2013; Bullock, 2008). Previously,
extracted audio features are obtained by using different type of audio feature extraction
techniques such as Mel-Frequency Cepstral Coefficient (MFCC), Linear Predictive Coding
(LPC) and Zero-Crossing Rate (ZCR) for musical instrument recognition describing various
sound qualities (Bormane & Dusane 2013; Umapathy, Krishnan & Rao, 2007). However,
according to Kumari, Kumar and Solanki (2010), the audio features become quite hard to
extract robustly when dealing with musical phrases such as bass-line, percussion loops and
others. Therefore, choosing the right feature extraction techniques to extract the features is
crucial. In addition, the field of music feature extraction is a wide research area, and
improving feature extraction will most likely have a major impact on the performance of an
instrument classification system (Gunasekaran & Revathy, 2008). In the next section, the
research will provide detail descriptions on the chosen audio feature extraction techniques.
2.3 An Overview of Audio Feature Extraction Techniques
There are many techniques that has been used in audio feature extraction such as as Mel-
Frequency Cepstral Coefficients (MFCC), Local Discriminant Bases (LDB), Linear
Predictive Coding (LPC), Perceptual Linear Prediction (PLP) and Zero-Crossing Rate (ZCR)
(Dave, 2013; Anusuya & Katti, 2011; Umapathy, Krishnan & Rao, 2007). In addition, these
techniques have their own strengths and drawbacks in extracting the audio signals (Dave,
2013; Keronen et al., 2011; Ngo, 2011; Patil et al., 2012; Sheetal & Raut, 2012; Umapathy,
Krishnan & Rao, 2007). Table 2.1 provides the comparison of audio feature extraction
techniques used by researchers. For example, audio feature extraction techniques such as
MFCC has proven to be one of the effective techniques used in audio feature extraction
which has given a better performance, especially in audio recognition rate (Kumar, Pandya &
Jawahar, 2014; Xie, Cao & He, 2012; Keronen et al., 2011; Furht, 2009). According to
Chachada & Kuo (2014), MFCC features have performed better than LDB features in the
domain of artificial (instrumental) and natural sounds. In another study conducted by Dave
(2013), the author used MFCC, LPC and PLP audio feature extraction techniques and have
stated that compared to MFCC and LPC, PLP is more useful in the domain of speech signals
since PLP can discard irrelevant information of speech audio signals whereby improve the
speech recognition rate. Furthermore, compared to LDB and PLP, audio feature extraction
10
techniques such as MFCC, LPC and ZCR particularly has been applied many times by the
previous researchers in the domain of music instrument, specifically in the traditional Indian
musical instrument (Chougule & Chavan, 2014; Bormane & Dusane, 2013; Anusuya & Katti,
2011; Kumari, Kumar & Solanki, 2010; Gunasekaran & Revathy, 2008). Even though,
MFCC, LPC and ZCR have shown their good performance, the previous researches have
pointed out the drawback of these three establish techniques in extracting noisy signals in the
highest amplitude which caused quality degradation in the extracted features and may lead to
low classification accuracy (Stulov & Kartofelev, 2014; Dave, 2013; Wolf & Nadeu, 2008;
Subramanian, 2006). Detail explanation on techniques is described in the next subsection.
Table 2.1: Comparison of audio feature extraction techniques used by authors
Author/
Year
Feature
extraction
techniques
Classifiers Data Result
Chachada
and Kuo
(2014)
MFCC
Local
Discriminant
Bases (LDB)
K-Nearest
Neighbor (kNN)
Support vector
machines (SVM)
37 classes of
artificial
(instrumental)
and natural
sounds
(Example:
Bells,
Clapping,
Thunder and
etc.)
MFCC features have performed
better than LDB features by
using kNN in the domain of
artificial (instrumental) and
natural sounds
Kumar,
Pandya and
Jawahar
(2014)
MFCC K-Nearest
Neighbor (kNN)
Bayes Network
(BNs)
Support Vector
Machines (SVM)
10 Indian
ragas
(melody)
consisting of
170 tunes
In recognition of raagas
(melody), kNN and BNs
achieved more than 70% of
classification accuracy and SVM
achieved more than 80% of
classification accuracy by using
MFCC
Bormane &
Dusane
(2013)
MFCC
LPC
ZCR
Wavelet Packet
Transform
(WPT)
4 classes of
musical
Instruments
(Example:
Sitar, Piano,
Guitar, and
etc.)
Family Notes
Recognition
String
Instruments
More than 60%
of recognition
rate
Keyboard
Instruments
More than 50%
of recognition
rate
Woodwind
Instruments
More than 50%
of recognition
rate
Brass
Instruments
More than 60%
of recognition
rate
11
Table 2.1 (continued)
Author/
Year
Feature
extraction
techniques
Classifiers Data Result
Dave
(2013)
MFCC
LPC
PLP
Support Vector
Machine (SVM)
Artificial Neural
network (ANN)
Music and
Speech
Signals
Discussion
Compared to MFCC and LPC,
PLP is more useful in the domain
of speech signals
ANN is more preferable to be used
in the field of speech recognition
Rocha,
Panda
and Paiva
(2013)
Standard
audio features
(SA) (spectral
features and
MFCC )
Melodic
audio features
(MA)
K-Nearest
Neighbor (kNN)
Bayes Network
(BNs)
Support Vector
Machines (SVM)
C4.5 decision tree
Naïve Bayes (NB)
903 datasets
of emotional
music
Classifier Best result
BNs 40% - 62%
NB 39% - 48%
kNN 40% -60%
C4.5 decision
tree
34% - 60%
SVM 45% - 64%
Patil
(2012)
MFCC
LPC
ZCR
Short-Time
Energy (STE)
Spectral
Feature ( SF)
Support Vector
Machines (SVM)
Humming
sounds
produced by
two male
speakers
The fusion system of
MFCC+ZCR+STE+SF gives
86.29% of classification accuracy
compared to the MFCC which is
77.71%.
Sheetal
and Raut
(2012)
LPC
Wavelet Packet
Transform (WPT)
Music and
Speech
Signals
Discussion
LPC can remove the redundancy
in the signal and has the highest
rate of audio compression
Xie, Cao
and He
(2012)
MFCC
LPC
ZCR
Support Vector
Machine (SVM)
40 samples
sound of
cutting
down trees,
sawing trees
and trees
collapse
Compared to LPC and ZCR,
MFCC can give better
performance in audio recognition
with more than 80% of recognition
rate.
Anusuya
and Katti
(2011)
MFCC
LPC
PLP
Wavelet Packet
Transform (WPT)
500 samples
of clean and
noisy
Kannada
audio
signals
Clean
signal
Noisy
signal
MFCC 70-90% 50-70%
LPC 70-80% 50-60%
PLP 40-60% 60-70%
12
Table 2.1 (continued)
Author/
Year
Feature
extraction
techniques
Classifiers Data Result
Keronen et
al.
(2011)
MFCC
Gaussian
Mixture Model
(GMM)
Recorded
audio signals
contained
noise from
public place
and car
environments
The performance of MFCC
degraded due to the mismatch
between the training and
recognition noise environments
Ngo
(2011)
LPC
ZCR
SVM Noisy audio
signal
captured from
the
microphone
and outputted
to the
loudspeaker
in hearing
aids
ZCR has high signal frequency
rate in unvoiced sound compared
to voiced sound
Kumari,
Kumar and
Solanki
(2010)
MFCC
Auto-
correlation
Artificial Neural
network (ANN)
5 different
type of North
Indian
musical
instruments
MFCC Autocor
-relation
Flute 61% 79.3%
Dholak 77.0% 80.7%
Sitar 60%
Bhapang 64.70%
Instrument
family
72 %
Gunasekaran
and Revathy
(2008)
MFCC
K-Nearest
Neighbor (kNN)
Multi-layer
perceptron
(MLP)
Gaussian
Mixture
Model(GMM)
10 Indian
musical
instrument
(Example:
Veena, Sitar,
Indian Flute
and etc.)
Confidence-based Fusion results
KNN + MLP 93.6%
KNN + GMM 92.8%
MLP+GMM 90.9%
KNN+MLP+GWM 92.1%
Umapathy,
Krishnan
and Rao
(2007)
Local
Discriminant
Bases (LDB)
MFCC
Local
Discriminant
Bases (LDB)
Artificial and
natural
sounds
(example:
drums, flute,
animals and
etc.)
(213 sounds)
LDB & MFCC (level of average
accuracy)
First level 91%
Second level 99%
Third level 95%
13
2.3.1 Mel Frequency Cepstral Coefficients (MFCCs)
Mel Frequency Cepstral Coefficients (MFCCs) are cepstral coefficients used for representing
audio in a way that mimics the physiological properties of the human auditory system
(Kumari, Kumar & Solanki, 2010; Sukor, 2012). The mel scale was developed based on the
study of human auditory perception. MFCCs are commonly used in speech recognition and
are increasingly used in music information recognition and genre classification systems
(Kumari, Kumar & Solanki, 2010). MFCC technique has its own advantages and
disadvantages in extracting audio signal. According to Xie, Cao and He (2012), MFCC will
give better performance in audio recognition rate. In addition, MFCC will take a short time
for extracting the features of audio signal (Keronen et al., 2011). However, the drawback of
MFCC is it will produce low quality of audio features and leads to low classification
accuracy if the signal used is noisy (Anusuya & Katti, 2011). Figure 2.1 shows the block
diagram of MFCC.
Figure 2.1: MFCC Block diagram (Dave, 2013)
From the MFCC block diagram shown in Figure 2.1, the audio signal from the Indian
musical instrument is passed into a process called pre-emphasis. The function of the pre-
emphasis process is to boost the energy of signal and amplify the importance of high-
frequency formant (Chougule & Chavan, 2014; Le-Qing, 2011; Zhu & Alwan, 2003).
Formants are the area in a spectrogram that shows the presence of noise in the highest
amplitude by displaying the area as dark bands. Besides, the darker formants produced in the
spectrogram shows the audio signals have a stronger energy (amplitude) (Prahallad, 2011;
Aslam et al., 2010). The function of the spectrogram is to plot the audio signal in amplitude,
frequency or time as shown in Figure 2.2.
Mel
Cepstrum Mel
Spectrum Spectrum
Frame
Pre-emphasis Frame
Blocking
DFT
Windowing
Cepstrum Mel-frequency
Wrapping
Audio Signal
14
Figure 2.2: Spectrogram diagram (McMurray, 2004)
The equation for pre-emphasis is shown in Equation (2.1):
H(z)=1-az-1 (2.1)
where z is the Discrete Fourier Transform (DFT) of input signal and a is the pre-emphasis
alpha coefficient where the value is usually between 0.9 and 1.0 (the standard value is 0.97).
Then, the signal will be passing to the frame blocking where the signal will be
blocked into frames of N samples, with adjacent frames being separated by M (M<N). The
first frame consists of first N samples; second frame begins with M samples after the first
frame, and overlaps it by N-M samples and so on (Gadade, Jadhav & Deogirkar, 2010). The
standard parameter or value for M and N is M = 100 and N = 256. The next step is windowing
where each individual frame is then windowed by using Hamming window in order to
minimize the signal discontinuities at the beginning and end of each frame by taper the signal
to zero (Dave, 2013). The Hamming window, w(n) is computed according to the Equation
(2.2):
w(n) = 0.54 – 0.46 cos (2πn / N-1), 0 ≤ n ≤ N -1 (2.2)
where N is total number of sample and n is current sample.
Then, the signal will be processed by using Discrete Fourier Transform (DFT). DFT
convert the sampled function from its original domain (often time or position along a line) to
the frequency domain. Therefore, each frame of N samples from the time domain is converted
15
into the frequency domain which is defined on the set of N samples {xn}, as shown in
Equation (2.3):
N-1
Xk = ∑ xne-j2πkn / N, k = 0, 1, 2,…, N-1 (2.3)
n=0
where Xk’s are complex numbers and their absolute values (frequency magnitudes) are
calculated. Xk is interpreted as follow: positive frequencies 0 ≤ f < Fs / 2 correspond to values
0 ≤ n ≤ N / 2-1, while negative frequencies -Fs / 2 < f <0 correspond to N / 2+1≤ n≤ N-1.
Here, Fs is considered as sampling frequency.
The output of DFT is defined in spectrum. In the next step, the spectrum is wrapped
through a process named mel-frequency wrapping and expressed in the mel frequency scale.
According to Dave (2013), mel-frequency scale is linear frequency spacing below 1000 Hz
and a logarithmic spacing above 1000 Hz. The output is defined as mel spectrum. The
approximate formula to compute the mel is shown in Equation (2.4):
mf = 2595log10 (f /700 + 1) (2.4)
where f is frequency and measured in Hz.
Cepstrum inverses the Fourier transform of the logarithm of the estimated spectrum of an
audio signal as shown in Equation (2.5):
K
Ĉn=∑(log Ŝk) cos[n*(k-0.5)*π/K], n=0,1, ..., K-1 (2.5)
k=1
where K is a number of mel spectrum coefficients.
From the equation (2.5), the result is expressed in mel cepstrum and is referred to as the mel-
scale cepstral coefficients, or MFCC which is shown in Equation (2.6):
16
M
Ci (n) = ∑ S (m) cos[πn(m-0.5)/M] , 0≤m<M (2.6) m-1
where n is the number of MFCC, Ci(n) is the n-th MFCC coefficients of the i-th frame, S(m)
is the logarithmic power spectrum of the audio signal, and M is the number of triangular
filters.
2.3.2 Linear Predictive Coding (LPC)
Similar to MFCC, Linear Predictive Coding (LPC) is another technique which offers a
powerful, yet simple method to extract audio information. LPC algorithm produces a vector
of coefficients that represents a smooth spectral envelope of a temporal input signal (Elminir,
ElSoud & El-Maged, 2012). The strength of LPC is it can remove the redundancy in signal.
In addition, LPC has the highest rate of audio compression (Sheetal & Raut, 2012) and take
short training time, even though it will take long time to extract the feature of audio signal
(Wolf & Nadeu, 2008). Nevertheless, using LPC alone for the recognition process is not very
successful because the all pole assumption of the vocal cord transfer function was not
accurate (Wolf & Nadeu, 2008). On the other hand, LPC coder does not work well for low or
high pitch frequency of voices (speech) signal (Kamal, Sarkar & Rahman, 2011). Another
importance drawback of LPC is the extracted audio features may lead to low performance of
audio classification accuracy if there is noise in the highest amplitude of an audio signal
(Chougule & Chavan 2014; Dave, 2013). This is due to LPC synthesizer process amplified
the noise in highest amplitude that may decrease the quality of extracted features. Figure 2.3
shows the LPC block diagram.
17
Figure 2.3: LPC Block diagram (Alwan, 2002)
First, the audio signal will be sent to LPC analyzer to determine the key features of
the signal and try to encode the signal as accurately as possible. The key features resemble
the loudness of the signal determine whether the sound is voiced or unvoiced. Then, the
signal is sent to the channel. The function of the channel is to transmit or serve as the medium
for transmission. Then, the signal will be passed to a decoder. Decoding involves using the
parameters acquired in the encoding and analysis to build a synthesized version of the
original audio signal. From the decoder, the signal is passed to LPC synthesizer. The function
of LPC synthesizer is as a computerized console or module for creating or scaling the audio
signal (Alwan, 2002). LPC synthesizer will scale the output signal to match the level of the
input signal (McCree, 2008). In other word, if the input signal contains high energy
(amplitude), LPC synthesizer will amplify the output signal to match with the input signal.
The formula for LPC synthesizer is shown in Equation (2.7):
p
s(n) = ∑αk s(n-k) + G u(n) (2.7) k=1
where s(n) is the waveform samples, αk is the predictor coefficient, and G represents the
loudness and it is multiplied with the excitation signal, u(n) to obtain proper loudness
intensity in the excitation signal.
From equation (2.7), the result can be simplified and referred as the Linear Predictive Coding
(LPC) which is shown in Equation (2.8) (Aviv & Grichman, 2011):
s_out(n)
s(n) LPC Analyzer Channel
LPC
Synthesizer
Decoder
18
p
H(z) = G / (1 + ∑ ap (k) z-k ) (2.8) k=1
where p is the order (number of poles), gain G is the signal loudness, and { ap(k) } are the
input parameters for the audio files.
2.3.3 Zero-Crossing Rate (ZCR)
Zero-crossing rate (ZCR) is the rate at which the signal changes from positive to negative or
vice versa. Simply, ZCR is a measure of the number of times in a given frame that the
amplitude of the audio signals passes through a value of zero. The rate at which zero crossing
occurs is a simple measure of the frequency content of a signal. Figure 2.4 shows the block
diagram of ZCR. First, the parameter of audio file such as frame size is assigned. Then, the
amplitude of the audio signal is calculated in energy calculation and the output is called ZCR
which is shown in Equation (2.9):
Figure 2.4: ZCR Block diagram (Raju et al., 2013)
T-1
ZCR = (1/T-1) ∑ || { st st -1<0 } (2.9) t=1
where, s is a signal of length T and the indicator function || {A} is equal to 1 when A is true
and is equal to 0 otherwise.
In order to use ZCR to distinguish unvoiced sounds from noise and the environment,
the waveform can be shifted before computing the ZCR. ZCR has high signal frequency rate
and is much lower for voiced sound as compared to unvoiced sound (Ngo, 2011; Patil et al.,
Input Signal Output Signal Assign Value Calculate
Energy
19
2012). ZCR is an important parameter for voiced or unvoiced classification (Khan, Bhaiya, &
Banchhor, 2012). Unfortunately, the drawback of ZCR is it unable to extract the best quality
of audio features due to the energy calculation process that amplify the noise in the highest
amplitude of an audio signal (Bormane & Dusane, 2013). Therefore, ZCR leads to
degradation of audio classification accuracy (Bormane & Dusane, 2013).
Nevertheless, some improvement can be done to MFCC, LPC and ZCR to improve the
performance of the techniques in audio feature extraction whereby contribute to high quality
of audio features and lead to a better audio classification accuracy. Therefore, this inspired
the use of an equalizer to be integrated with audio feature extraction techniques to equalize
the noise in the highest amplitude of audio signal. The next section will provide a detail
explanation on the noise in audio signal.
2.4 Noise in Audio Signal
Noise can be defined as uncontrolled, loud, unmusical, or unwanted component of an audio
signal (Gold, Morgan & Ellis, 2011; Weeks, 2010; Thorne, 2007). In addition, there are many
types of noise such as voiced noise, unvoiced noise or background noise (Gold, Morgan &
Ellis, 2011). Voiced noise is the sound produced by the human voice such as breathing or
coughing, meanwhile unvoiced noise is the sound produced from the music instrument or
sound effects such as door knocks, paper shuffling or plucking a string instrument (Gold,
Morgan & Ellis, 2011; Gunasekaran & Revathy, 2008). Background noise is any sound other
than the sound being used than can be based on the surrounding or can be considered as an
external sounds such as ambulance going by outside or people talking (Gold, Morgan & Ellis,
2011). In this research, the noise has been identified come from the music instrument itself
due to the vibration of the string that produced noise in the highest amplitude (Dave, 2013;
Stulov & Kartofelev, 2014; Subramanian, 2006). Therefore, some type of noise such as
background noise could be reduced or removed by using noise filtering. Specifically, some
type of noise such as high amplitude noise can only be mitigated by using sound equalizer
(Kaur & Kansal, 2013; Gold, Morgan & Ellis, 2011; Weeks, 2010). The next section will
describe on the noise filtering and sound equalizer in more detail.
20
2.5 Noise Filtering
Noise filtering or also known as noise reduction is a process of removing unwanted
components (noise) from an audio signal (Tan & Jiang, 2013). Noise filtering is normally
applied in audio signal to perform application such as noise reduction, sound crossover and
others (Tan & Jiang, 2013; Altera, 2010). Basically, noise filtering involves the process of
removing different types of noise such as background noise or unvoiced noise from the audio
signal (Gold, Morgan & Ellis, 2011; Weeks, 2010). Moreover, noise filtering is used to filter
a noisy signal, such as cleaning up an audio signal recorded in a room full of other
conversations (Tan & Jiang, 2013). Nevertheless, noise filtering does not focusing on the
specific high amplitude noise while filtering background noise in the audio signal (Tan &
Jiang, 2013; Weeks, 2010). Therefore, as an alternative to the drawback of noise filtering,
this study suggests the use of sound equalizer. The sound equalizer has the potential of
equalizing the noise that come from the high amplitude of an audio signal (Kaur & Kansal,
2013; Kumar & Kaur, 2012; Altera, 2010). The next section will provide a detail explanation
on the sound equalizer.
2.6 Sound Equalizer
Sound equalizer is an essential part of any sound system which provides an approximate
inverse of the channel frequency response (Kumar & Kaur, 2012). Equalizers are used in
recording studios, broadcast studios, and live sound reinforcement to eliminate unwanted
sounds such as noise from microphones, instrument pick-ups, loudspeakers, and hall
acoustics. Figure 2.5 shows an audio equalizer in the communication system. In general, the
audio captured from the audio input source is sent to an equalizer by using a communication
channel (physical medium) such as wires, radio, acoustic, magnetic or optical recording
media. The function of an equalizer is to equalize the amplitude of an audio signal and
transmit to a digital to analog converter (DAC). DAC will convert the digital audio signal to
analog sound and output the analog sound to a speaker or headphone.
21
Figure 2.5: Audio equalizer (Proakis, 2008)
Specifically, equalizer involves the process of equalization to mitigate the effects of
intersymbol interference (ISI). ISI is a form of distortion of audio signal due to an unwanted
sound or effect such as noise which is produced in the high amplitude (Kaur & Kansal, 2013).
Kumar and Kaur (2012) also stated that the reduction of ISI effects has to be stabilized since
the audio signal contains noise. Based on Deepa (2013), ISI arises because of the spreading
of a transmitted pulse due to the dispersive (widely spread or scattered) nature of the channel,
which results in overlapping of adjacent pulses. Equalizer is usually implemented at the
original frequency or the amplitude range of the signal (National Chung Cheng University,
2013). Equalization equation is shown in (2.10):
y(t) = x(t) * f*(t) + nb(t) (2.10)
where
x(t) is the original input signal
f*(t) is the complex conjugate of f(t)
(Changing the sign of the imaginary part and leaving left real part unchanged)
nb(t) is the noise at the input of the equalizer
Equalizer responds to the impulse response of the signal. Impulse response refers to
the reaction of wave in signal response to some external change (Golden, Dedieu & Jacobsen,
2005). High wave in signal refers to the highest amplitude of wave in the channel. The
simplest way to obtain the channel impulse response is to send a single pulse (approximately
an impulse) over the channel and observe the resulting received signal. If there is no noise on
the channel, this could give a good estimate of the channel impulse response. However, in
reality, an estimate of the channel impulse response that is based on a single transmitted
Input
In Out
In Out
In Out
Output
Audio
capture
Communication
channel Equalizer Write to
DAC Audio
Playback
22
impulse will always be noisy. Therefore, the possibility to reduce the effect of the noise is by
taking many measurements of the impulse response (Golden, Dedieu & Jacobsen, 2005).
There are two general categories of equalizer which is linear and nonlinear equalizer.
Altera (2010) stated that linear equalizer is frequency dependent and it can be highly effective
in mitigating the ISI. In summation, a linear equalizer mitigates the ISI of a single audio
signal without enhancing the noise. Some examples of linear equalizer are Zero-Forcing
Equalizer (ZFE) and Minimum Mean Square Error (MMSE) equalizer. On the other hand,
nonlinear equalizer involves the use of mixed-signal (more than one audio signal in the
waveform). MIT Lincoln Laboratory (2009) stated that nonlinear equalizer is used to detect
small signals in the presence of strong background, such as in radar, signal intelligence, and
electronic intelligence systems. According to National Chung Cheng University (2013), some
of the example of nonlinear equalizer is Maximum Likelihood Symbol Detection (MLSD)
and Maximum Likelihood Sequence Estimator (MLSE). This study will only focusing on
linear equalizer that are ZFE and MMSE since linear equalizer highly effective in mitigating
the unwanted sound (noise) in the highest amplitude of audio signal (Kaur & Kansal, 2013;
Kumar & Kaur, 2012; Altera, 2010). The next subsection will discuss ZFE and MMSE in
detail.
2.6.1 Zero Forcing Equalizer (ZFE)
Zero Forcing Equalizer is a linear receiver used in communication systems. This equalizer
inverts the frequency response of the channel. Kumar and Kaur (2012) stated that for a
channel with frequency response F(f), the zero forcing equalizer C(f) is constructed as shown
in Equation (2.11):
C(f) = 1 / F(f) (2.11)
The output of ZFE is calculated from the numerator and denominator coefficient, which is
obtained from the inverse of impulse response in the frequency domain, F(f). Thus, the
combination of channel and equalizer gives a flat frequency response and linear phase as
shown in Equation (2.12):
F(f)C(f) = 1 (2.12)
23
The implementation of ZFE depends on the channel it is used (Proakis, 2008). A channel is
used to convey an information signal, for example a digital bit stream, from one or several
senders (or transmitters) to one or several receivers. In information theory, a storage device is
also a kind of channel, which can be sent to (written) and received from (read). A typical
channel is model in discrete domain as shown in Equation (2.13) (Dytso, 2012):
y[n] = h[n]*x[n] + z[n] (2.13)
where
y[n] is the channel output
h[n] is the channel impulse response
x[n] is the channel input
z[n] is the noise
From equation (2.13), the channel is converted to frequency domain as shown in Equation
(2.14):
y(f) = h(f)x(f) + z(f) (2.14)
ZFE multiplies y(f) and z(f) by inv(h(f)) to reduce ISI as shown in Equation (2.15) (Kaur &
Kansal, 2013):
inv(h(f))y(f) = x(f) + inv(h(f))z(f) (2.15)
In previous research, ZFE has been used to solve the problem of signal transaction in
Multiple Input Multiple Output (MIMO) systems (Khademi et al., 2013). MIMO technology
is a wireless technology that uses multiple transmitter and receiver to improve
communication performance. Hence, ZFE is used to mitigate the interference in signal
transaction. Moreover, ZFE is much more useful for equalizing the effect of noise in the
higher amplitude as introduced by ISI (Kaur & Kansal, 2013; Mobile Communication, 2009).
However, the drawback of ZFE is that the channel response may often exhibit attenuation
(reduction of signal strength during transmission) at high frequencies around one-half the
sampling rate (the folding frequency) (Kumar & Kaur, 2012). Another drawback is that the
use of the equalizer as standalone or independently decreases the performance of the channel.
Therefore, it depends on how ZFE is used as mentioned by Dytso (2012).
24
As mentioned previously, due to the drawback of audio feature extraction techniques
in extracting noisy signal, ZFE is proposed to be integrated with audio feature extraction
techniques since the characteristics of ZFE is to mitigate the effect of noise in the highest
amplitude of the audio signal. Hence, ZFE will be able to improve the performance of audio
feature extraction techniques. Another equalizer known as MMSE, which is in the same
category with ZFE is discussed in the next section.
2.6.2 Minimum Mean Square Error (MMSE) Equalizer
Minimum mean square error (MMSE) equalizer minimizes the mean square error (MSE).
MSE is a common measure of estimator quality as stated by Kumar and Kaur (2012). The
main function of MMSE equalizer is that it does not usually eliminate ISI completely but it
minimizes the total power of the noise in the output. If x is an unknown random variable, then
an estimator of x will be any function from the measurement of known random variable, and
its MSE is given by the trace of error as shown in the simplified equation (2.16):
MSE =E { ( -x )2 } (2.16)
where x is a scalar variable.
According to Cioffi (2008), MMSE can provide better performance if the audio signal
is voice (speech). However, the author pointed out the major drawback of MMSE is that the
equalizer is slightly more complicated to describe and analyze than the ZFE. Also, because of
the biasing (there is an external force that controls the equalizer), the MMSE output is
slightly lower than the ZFE output. MMSE does not assume any stochastic mechanism
(having random variable) of the desired and observed signals (Chen et al., 2013). It only
makes assumptions about the noise. For example, the noise is additive zero-mean, time-
independent, bounded (limited), and known variance. It also does not usually reduce the ISI
effect. Due to the disadvantages of the MMSE, this research is focusing on ZFE. However,
this research will also integrate MMSE with audio feature extraction techniques in order to
compare the performance of both equalizers when they are integrated with the respective
audio feature extraction techniques.
71
REFERENCES
Agarwal, P., Karnick, H., & Raj, B. (2013). A Comparative Study Of Indian And Western Music
Forms. In ISMIR. Brazil: Pontifical Catholic University of Paraná. pp. 29-34.
Altera. (2010). Understanding the Pre-Emphasis and Linear Equalization Features in Stratix
IV GX Devices. Retrieved on January 29, 2014, from www.altera.com
Alwan, A. (2002). Wideband Speech Coding with Linear Predictive Coding (LPC). Retrieved
September 13, 2013, from http://www.seas.ucla.edu
Anusuya, M.A. & Katti, S.K. (2011). Comparison of Different Speech Feature Extraction
Techniques with and without Wavelet Transform to Kannada Speech Recognition,
International Journal of Computer Applications, 26(4), pp. 19-24.
Aslam, T., Doherty, J., Dubrau, A., & Hendren, L. (2010). AspectMatlab: an aspect-oriented
scientific programming language. In Proceedings of the 9th International Conference
on Aspect-Oriented Software Development. France: ACM New York. pp. 181-192.
Atlas, L. (2011). Spectrogram. Retrieved on March 18, 2015, from http://isdl.ee.washington.edu
Aviv, A., Grichman, K. (2011). Long-term prediction. Retrieved on August 30, 2013, from
http://health.tau.ac.il
72
Bello, J. P. (2013). Sound Classification. Retrieved on May 2, 2014, from http://www.nyu.edu
Ben Gal, I. (2007). Bayesian Networks, in Encyclopedia of Statistics in Quality and Reliability,
Ruggeri, F., Kenett, R. S. and Faltin, F. (editors in chief). USA: John Wiley & Sons Ltd.
Bormane, D. S., & Dusane, M. (2013). A Novel Techniques for Classification of Musical
Instruments. In Information and Knowledge Management, 3(10), pp. 1-8.
Buckingham Music Inc. (2008). Veenas. Retrieved on February 4, 2013, from
http://www.buckinghammusic.com/veena/veena.html
Bullock, J. (2008). Implementing audio feature extraction in live electronic music. University
of Birmingham: Ph.D. Thesis.
Chachada, S., & Kuo, C. C. J. (2014). Environmental sound recognition: A survey. APSIPA
Transactions on Signal and Information Processing, 3(14), pp. 1-15.
Chen, J., Ma, T., Chen, W., & Peng, Z. (2013). Unsupervised robust recursive least-squares
algorithm for impulsive noise filtering. Journal Science China Information Sciences,
56(4), pp. 1-10.
Cioffi, M.J. (2008). Chapter 3. Equalization. Retrieved on January 30, 2014, from
http://www.stanford.edu/group/cioffi/book/
Chougule, S. V., & Chavan, M. S. (2014). Channel Robust MFCCs for Continuous Speech
Speaker Recognition. Advances in Signal Processing and Intelligent Recognition
Systems, pp. 557-568.
73
Chu, S., Narayanan, S., & Kuo, C. C. (2009). Environmental sound recognition with time–
frequency audio features. IEEE Transactions on Audio, Speech, and Language
Processing, 17(6), pp. 1142-1158.
Courtney, D. (2012). Gottuvadhyam. Retrieved on February 3, 2013, from
http://chandrakantha.com
Dave, N. (2013). Feature Extraction Methods LPC, PLP and MFCC in Speech Recognition.
International Journal of Research in Engineering and Advanced Technology, 1(6),
pp. 1-5.
Deepa,T. (2013). Linear Equalizers & Nonlinear Equalizers. Retrieved on January 28, 2014,
from http://www.srmuniv.ac.in
Dytso, A. (2012). Zero Forcing Equalizer. Retrieved on January 29, 2014, from
http://www.mathworks.com
Elminir, H. K., ElSoud, M. A., & El-Maged, L. A. (2012). Evaluation of Different Feature
Extraction Techniques for Continuous Speech Recognition. International Journal
of Science and Technology, 2(10), pp. 689-695.
Eronen, A. (2001). Automatic musical instrument recognition. Tempere University of
Technology: Master’s Thesis.
Furht, B. (2009). Handbook of multimedia for digital entertainment and arts. USA: Springer.
74
Gadade, M. H., Jadhav, M. M. R., & Deogirkar, M. S. V. (2010). Speech Identification and
Recognition Using Data Mining. Government College of Engineering: Technical
Report.
Gold, B., Morgan, N., & Ellis, D. (2011). Speech and audio signal processing: processing and
perception of speech and music. USA: John Wiley & Sons.
Golden, P., Dedieu, H. & Jacobsen, K.S. (2005). Fundamentals of DSL Technology. USA:
Auerbach.
Gunasekaran, S. & Revathy, K. (2008). Fractal dimension analysis of audio signals for Indian
musical instrument recognition. International Conference on Audio, Language and
Image Processing (ICALIP). Shanghai: IEEE. pp. 257-261.
Gupta, S., Jaafar, J., Ahmad, W. F., & Bansal, A. (2013). Feature Extraction Using MFCC. An
International Journal of Signal & Image Processing (SIPIJ), 4(4), pp. 101-108.
Jupiter Infomedia Ltd. (2012). Carnatic music. Retrieved on December 24, 2012, from
http://www.indianetzone.com/2/carnatic_music.htm
Jupiter Infomedia Ltd. (2011).Veena. Retrieved on December 24, 2012, from
http://www.indianetzone.com/16/veena.htm
Kamal, S. N., Sarkar, R. N., & Rahman, M. (2011). Comparison of Musical Pitch Analysis
Between LPC and CELP. International Journal of Advances in Engineering Sciences,
1(1), pp. 35-39.
75
Kaur, N., & Kansal, L. (2013). Performance Comparison of MIMO Systems over AWGN and
Rician Channels with Zero Forcing Receivers. International Journal of Wireless &
Mobile Networks, 5(1), pp. 73-84.
Keronen, S., Pohjalainen, J., Alku, P., & Kurimo, M. (2011). Noise robust feature extraction
based on extended weighted linear prediction in LVCSR. In Proceedings of
INTERSPEECH, Italy: ISCA. pp. 1265-1268.
Khademi, S., Chepuri, S. P., Leus, G., & van der Veen, A. J. (2013). Zero-forcing pre-
equalization with transmit antenna selection in MIMO systems. International Conference
on Acoustics, Speech and Signal Processing (ICASSP), Vancouver: IEEE. pp. 5046-5050.
Krishnaraj S. (2010). Carnatic music. Retrieved on January 7, 2013, from
http://www.carnaticindia.com/
Kumar, V., Pandya, H., & Jawahar, C. V. (2014). Identifying ragas in Indian music. 22nd
International Conference on Pattern Recognition (ICPR). Sweden: IEEE. pp. 767-772
Kumar, M., & Kaur, J. (2012). Performance analysis of BPSK system with ZF &MMSE
equalization. International Journal of Latest Trends in Engineering and Technology
(IJLTET).1(3), pp. 127-133.
Kumari, M., Kumar, P. & Solanki, S. S. (2010). Classification of North Indian Musical
Instruments using. Spectral Features. Journal of Computer Science and
Telecommunications (GESJ). 6(29), pp. 11-24.
76
Le-Qing, Z. H. U. (2011). Insect sound recognition based on MFCC and PNN. International
Conference on Multimedia and Signal Processing (CMSP). China: IEEE. pp. 42-46.
Lerch, A. (2012). An introduction to audio content analysis: Applications in signal processing
and music informatics. USA: John Wiley & Sons.
Li, T., Ogihara, M., & Li, Q. (2003). A comparative study on content-based music genre
classification. In Proceedings of the 26th annual international ACM SIGIR conference on
Research and development in information retrieval. Canada: ACM New York. pp. 282-
289.
McCree, A. (2008). Springer Handbook of Speech Processing. India: CBS. pp. 97-105.
McMurray, B. (2004). Experimental Psychology II. Retrieved on August 20, 2014, from
http://www2.psychology.uiowa.edu/
MIT Lincoln Laboratory. (2009). Nonlinear Equalization for Receiver Dynamic Range
Extension. Massachusetts Institute of Technology: Technical Report.
Mobile Communication (2009). Equalization, Diversity and Coding Techniques. Retrieved on
August 30, 2013, from http://mc.lctu.cn
Nasridinov, A., & Park, Y. H. (2014). A Study on Music Genre Recognition and Classification
Techniques. International Journal of Multimedia and Ubiquitous Engineering. 9(4), pp.
31-42.
National Chung Cheng University. (2013). Equalization, Diversity, and Channel Coding.
Retrieved on January 28, 2014, from web.ee.ccu.edu.tw
77
Nettleton, D. F., Orriols-Puig, A., & Fornells, A. (2010). A study of the effect of different types
of noise on the precision of supervised learning techniques. International Journal of
Artificial intelligence review, 33(4), pp. 275-306.
Ngo, K. (2011). Digital signal processing algorithms for noise reduction, dynamic range
compression, and feedback cancellation in hearing aids. University of Leuven: Ph.D.
Thesis.
Panda, A. K., & Sahoo, A. K. (2011). Study of Speaker Recognition Systems. Georgia Institute of
Technology: Ph.D. Thesis.
Peng, L., Yang, D., & Chen, X. (2014). Multi frame size feature extraction for acoustic event
detection. In Asia-Pacific Signal and Information Processing Association, 2014 Annual
Summit and Conference (APSIPA). Cambodia: IEEE. pp. 1-4.
Patil, H. A., Madhavi, M. C., Jain, R., & Jain, A. K. (2012). Combining Evidence from Temporal
and Spectral Features for Person Recognition Using Humming. Proceedings of the First
Indo-Japan conference on Perception and Machine Intelligence. India: Springer. pp. 321-
328.
Prahallad, K. (2011). Speech Technology. Retrieved on April 27, 2014, from
http://www.docstoc.com/docs/109146955/Topic-Spectrogram_Cepstrum-and-Mel-
Frequency-Analysis
Proakis J, Salehi M (2008). Digital Communications. 5th edition. New York: McGraw-Hill.
78
Rahman, R. M., & Afroz, F. (2013). Comparison of Various Classification Techniques Using
Different Data Mining Tools for Diabetes Diagnosis. Journal of Software Engineering
and Applications, 6(03), pp. 85-97.
Raju, N., Arjun, N., Manoj, S., Kabilan, K., & Shivaprakaash, K. (2013). Obedient Robot with
Tamil Mother Tongue. Journal of Artificial Intelligence, 6(2), pp. 161-167.
Rao, B. T., Chinnam, S., Kanth, P. L., & Gargi, M. (2012). Automatic Melakarta Raaga
Identification Syste: Carnatic Music. International Journal of Advanced Research in
Artificial Intelligence (IJARAI), 1(4), pp. 43-48.
Rao, V., Gupta, C., & Rao, P. (2013). Context-aware features for singing voice detection in
polyphonic music. In Adaptive Multimedia Retrieval. Large-Scale Multimedia Retrieval
and Evaluation. Spain: Springer. pp. 43-57.
Rocha, B., Panda, R., & Paiva, R. P. (2013). Music Emotion Recognition: The Importance of
Melodic Features. 6th International Workshop on Machine Learning and Music. Czech
Republic: Springer. pp. 1-4.
Sathish, V M. (2010). South Indian star turns musician for Sharjah show. Retrieved on
February 3, 2013, from http://www.emirates247.com
Sheetal, G. D., & Raut, R. D. (2012). Advance Source Coding Techniques for Audio/Speech
Signal: A Survey. International Journal of Computer Technology & Applications, 3(4),
pp. 1335-1342.
79
Smith, S. (2013). Digital Signal Processing: A Practical Guide for Engineers and Scientists: A
Practical Guide for Engineers and Scientists. USA: Newnes.
SoundCloud (2007). Retrieved January 7, 2013, from https://soundcloud.com/
Stulov, A., & Kartofelev, D. (2014).Vibration of strings with nonlinear supports. Journal of
Applied Acoustics, 76(14), pp. 223-229.
Subramanian, M. (2006). Carnatic Music and the Computer. Retrieved on March 12, 2014, from
http://www.musicresearch.in
Sukor, A. S. (2012). Speaker identification using MFCC procedure and noise reduction method.
Universiti Tun Hussein Onn Malaysia: Master’s Project Report.
Tan, L., & Jiang, J. (2013). Digital signal processing: fundamentals and applications. USA:
Academic Press.
The Incredible India Travel (2011). Musical Instruments. Retrieved on January 5, 2013,
from http://www.theincredibleindiatravel.com
Thorne, R. (2007). Assessing intrusive noise and low amplitude sound. Massey University: Ph.D.
Thesis.
Umapathy, K., Ghoraani, B., & Krishnan, S. (2010). Audio signal processing using time-
frequency approaches: coding, classification, fingerprinting, and watermarking.
EURASIP Journal on Advances in Signal Processing, 10(1), pp. 1-28.
80
Umapathy, K., Krishnan, S. & Rao, R. K. (2007). Audio Signal Feature Extraction and
Classification Using Local Discriminant Bases. IEEE Transactions on Audio, Speech &
Language Processing, 15(4), pp. 1236-1246.
Walter, S., Scherer, S., Schels, M., Glodek, M., Hrabal, D., Schmidt, M., & Schwenker, F.
(2011). Multimodal emotion classification in naturalistic user behavior. In Human-
Computer Interaction. Towards Mobile and Intelligent Interaction Environments (pp.
603-611). Springer Berlin Heidelberg.
Weeks, M. (2010). Digital Signal Processing Using MATLAB and Wavelets. 2nd ed. USA: Jones
and Bartlett.
WEKA (2010). Retrieved February 3, 2013, from http://www.cs.waikato.ac.nz/ml/weka/
Witten, I., Frank, E., & Hall, M. (2011). Data Mining: Practical Machine Learning Tools and
Techniques. 3th edition. USA: Morgan Kaufmann.
Wolf, M. & Nadeu, C. (2008). Evaluation of different feature extraction methods for speech
recognition in car environment. 15th International Conference on Systems, Signals and
Image Processing (IWSSIP), Bratislava: IEEE. pp. 359-362.
Xie, C., Cao, X., & He, L. (2012). Algorithm of Abnormal Audio Recognition Based on
Improved MFCC. International Workshop on Information and Electronics Engineering
(IWIEE), China: Elsevier. pp. 731-737.
81
Yu, Y., Zimmermann, R., Wang, Y., & Oria, V. (2013). Scalable Content-Based Music Retrieval
Using Chord Progression Histogram and Tree-Structure LSH. IEEE Transactions on
Multimedia, 15(8), pp. 1969-1981.
Zhu, Q., & Alwan, A. (2003). Non-linear feature extraction for robust speech recognition in
stationary and non-stationary noise. Journal of Computer Speech & Language, 17(4), pp.
381-402.
Zytrax (2014). Digital Sound Primer. Retrieved on August 27, 2014, from
http://www.zytrax.com
